On the evening of July 16, 2019, Elon Musk and the Neuralink team organized a conference in San Francisco to present the progress of the work of the startup created in 2017. To better understand Musk’s last innovation, it was necessary to skin the video of the entire conference, the publication An integrated brain-machine interface platform with thousands of channels by Elon Musk and Neuralink, released in 2019 (12 pages), and the associated press coverage in the US tech press.
See in particular Elon Musk unveils Neuralink’s plans for brain-reading ‘threads’ and a robot to insert them – Not for humans, yet by Elizabeth Lopatto in TheVerge, Elon Musk’s Neuralink looks to begin outfitting human brains with faster input and output starting next year by Darrell Etherington in TechCrunch and Elon Musk’s Neuralink hopes to put sensors in human brains next year by Richard Lawler in Engadget. This allowed us to consolidate a maximum of information and illustrations for this article.
Neuralink’s presentation revolved around a technical solution of electrodes based on flexible wires implanted in the brain by a robot and controlled by an integrated circuit in a tiny box that can be integrated under the skull.
The first applications always target primarily the repair of a man suffering from various pathologies such as Parkinson’s disease, which is already treated with deep stimulation electrodes implanted in the subthalamic nucleus, motor control of artificial limbs.
For amputees via implantation of electrodes around the motor cortex at the lateral periphery of the brain, epilepsy, dystonia (involuntary muscle contractions), depressions, tinnitus, and finally, restoration of sight for the blind or hearing for the deaf. Such electrodes should also potentially allow disabled people to control a smartphone via related mechanisms associated with the motor cortex. This is the inventory provided by Neuralink.
Each use case will have its particularities: number of neurons to connect, their depth in the brain which can complicate the task, and their precise mapping for successful implantations. Not all these considerations were covered in Neuralink’s talk.
Neuralink is getting closer
Neuralink is not yet in the commercialization stage of its solution. They have so far tested it on rats and monkeys. They aim to obtain FDA approval to test it on humans by the end of 2020, in partnership with Stanford University to treat quadriplegic patients.
The next step would be to “increase” the abilities of people without disabilities, by connecting the brain directly to AI. It is, of course, a path much more distant in time, not particularly documented, and more uncertain. And Elon Musk wants to reassure us: “This is not a mandatory thing; this is something you can choose to have if you want”. Oh, God! It’s not reassuring at all, even presented in this light.
Beyond the classic media interest, the conference had another objective: facilitating recruitment tasks. Until now, the startup has secured $158M in funding, two-thirds from Elon Musk and it has 90 employees. Its president is Max Hodak.
Even if the project it’s still in an early stage, it is still worth exploring the four technologies presented by Neuralink: electrodes integrated with flexible wires, a robot to implant them, a chipset to control them, and software to operate them.
The connection with the brain passes through electrodes which are used either to measure the state of excitation of neurons by evaluating the electrical flows passing through them or by acting on them by electrical stimulation. The main novelty of Neuralink is miniaturized electrodes embedded in flexible wires. These threads are a quarter of the thickness of a hair and the size of a neuron.
Verification made: the diameter of a hair ranges from 17 to 181 μm depending on the case (source), the nucleus of a neuron has a size varying between 4 and 100 μm and the thickness of the Neuralink threads is between 4 μm and 6 μm. In comparison, the diameter of a red blood cell is 6 μm.
They created a 96-wire prototype comprising 3072 electrodes, so 32 electrodes per wire. They had two forms of wire: a linear version (A) or a tree (B), see below.
They have imagined twenty models that correspond to different geometries of biological neural networks to be targeted, and of which here are a few examples.
How are they made
These electrode wires are made with lithography techniques like those of CMOS semiconductor components. Neuralink even tested a process for manufacturing the chip and electrode wires simultaneously on a single wafer.
The advantage of these flexible electrodes is that they can track movement without damaging the internal brain. A soft body in a soft body is less dangerous than a hard body in a soft body. These electrodes would therefore be more suitable than those from Brain Gate, which were themselves produced by Cybernetics at the University of Utah.
They have been tested since 2002 in the USA, notably at Stanford, and aim to restore control of limbs. They allow the implantation of up to 128 electrodes.
Neuralink therefore mainly wants to increase the magnitude of the number of controllable neurons. The current model of 3,072 electrodes tested on rats could grow to 10,000 electrodes, which in itself is not a crazy progression.
Parkinson’s could be treated with them
Anyway, this technique would be better than the current stimulation systems for patients with Parkinson’s disease, which only includes about ten electrodes.
Of course, there are some reservations about this electrode technology. They seem adapted to the stimulation of cortical columns of the cortex but less adapted to the different parts of the limbic and deep brain, where Parkinson’s disease is treated.
Scientists are also wondering if their structure can resist the biological saline environment of the brain. Neuralink explained how it protects the processor but not the electrode leads.
Finally, these electrodes only measure the activity of neurons “in bundles” and without great precision. The measurement does not take place at all levels of connections between neurons, let’s say, for example, the famous synapses. This is sufficient, however, for the intended applications in the field of health.
The second technology presented by Neuralink is a large machine capable of implanting neurons with precision in the brain. Its design is not very detailed. In the demonstration, the robot implants wire electrodes on a rat’s cerebral cortex after drilling holes in the skull. It implants 6 wires per minute, containing 192 electrodes.
It does not implant these electrodes in the middle of the brain, obviously lacking an imaging system to navigate the brain “by sight”. The robot has so far been used on 19 animals with an 87% success rate, or 2.47 animals in which it has not worked.
Neuralink believes that one day they will be able to use a laser to punch holes in the cranium instead of using a mini drill. The idea is to make this kind of operation as mundane as LASIK, the use of a laser to treat myopia. The goal is also to eliminate the need for general anesthesia for implantation. But this is a promise that only binds those who hold it because the technical solution remains to be found.
Are his promises real?
Elon Musk often behaves like Elisabeth Holmes, former CEO of the failed company Theranos: he sells first and tries to do second, and he succeeds… or not. Two years ago, for example, he indicated that his electrodes would circulate in the blood to go to the right place in the brain!
Human implantation tests are expected to start by the end of 2020, with neurosurgeons at Stanford University. Once FDA gives its approval, a process that typically lasts at least a year, we will therefore see if it takes place before the marketing of the 5 autonomous Tesla which is also planned before the end of 2020.
Seen up close, the robot is like a sewing machine. The flexible wire implantation technique is vaguely reminiscent of that commonly used in Abbott’s Freestyle blood glucose sensors. The installation device is a system that projects a needle into the skin, in the middle of which is a wire for the electrochemistry of the interstitial fluid under the skin.
The needle goes under the skin with the thread and then retracts, leaving the thread in the skin. The needle remains attached to a small, outdoor coin-sized sensor. The manipulation performed by Neuralink’s robot looks a bit more complex and controlled by different cameras, under the supervision of a neurosurgeon.
The third component of the system: electrodes control box. The latest prototype, tested on rats, contains 12 ASICs (specialized processors) each capable of processing 256 electrodes, or the 3072 of the 96 wires that come out. It is connected externally via a USB-C port. These ASICs are fairly simple: they contain analog / digital amplifiers and converters. They call it an “analog pixel”.
ASIC can detect neural impulses which provide a means of compressing the large amount of information captured by the electrodes
The sample information at a frequency of 18.6 kHz with 10-bit resolution. ASIC consumes only 750 mW. For comparison, a smartphone consumes about 3 W. But this power is indicated for a scenario of recording neural states, not for their activation.
The size of the test case is 2.3cm x 1.85cm x 2mm thick. The goal is to miniaturize all this on a 4 x 5 mm chip integrated into airtight packaging to be grafted into the head.
Designed to live inside you
These chips are designed to be implanted in the human body. They will be connected directly to a kind of comb of electrodes to be implanted in the brain by a robot. They will then be connected by a cable to a flat solenoid installed inside the head behind the ear.
This will allow communication with a receiving device equipped with a battery that fits on the ear, the Link, via an unspecified wireless transmission (Bluetooth, Wi-Fi?). The solenoid must be able to be used to transmit power to the chips, such as contactless recharging your smartphone in Qi Power.
Neuralink teams indicated that this architecture made it possible to easily “unplug” the system, a way to calm fears of control of the human brain by an external machine. This should be put into perspective when you know how easy it is to unplug from your smartphone in everyday life!
Neuralink plans to first implant up to four of these chips: three in the motor cortex and one in the somatosensory cortex, which is not far from it. Everything is then controlled by a smartphone app.
Software and algorithms
This is the least elaborate and documented part of Neuralink’s lecture. They presented neural activity diagrams obtained from laboratory rats. Capture tests had been performed in several areas of the cerebral cortex of these rats. We can even see how the threads are implanted in the cortex of the brain of these unfortunate rats.
Ultimately, Neuralink should develop mobile applications that will allow us to read our minds to avoid going through the keyboard to enter a text or a command to launch an action. This is far from truly being an “AI affair.” There is no connection with any intelligence since it is just a means of entering individual letters. But don’t worry, the electrodes do not allow the results of your web searches to be implanted in your memory!
They also envision much more sophisticated actions to control our moods, help us solve math problems and other increases in intelligence, but that seems quite far-fetched and very speculative. For example, to control our moods, we should place their electrodes precisely in the middle of the brain, in the different areas triggering the production of hormones such as those of pleasure, fear, or satiety (thalamus, hypothalamus, hippocampus, amygdala,…).
They don’t do it yet. This may not necessarily be impossible to achieve, but it is much more complicated than placing electrodes on the cortex on the outskirts of the brain. I also think, as in 2017, that this would be both the most questionable and feasible application of these electrodes. Initially, we would control depression and other PTSD (Post-Traumatic Stress Disorder) and in the end, we would control moods! On the day of a vote, it might be useful!